14 May 2020, 07:00

Meet the winners of ICLR Workshop Challenge #2: Radiant Earth

Zindi is excited to introduce the winners of the ICLR Workshop Challenge #2: Radiant Earth Computer Vision for Crop Detection from Satellite Imagery Challenge. The challenge attracted 460 data scientists from across the continent and around the world, with 99 placing on the leaderboard.

The objective of this competition was to create a machine learning model to classify fields by crop type from images collected by the Sentinel-2 satellite during the growing season. The fields pictured in this training set are across western Kenya, and the images were collected by the PlantVillage team.

In contrast with a survey, agricultural maps based on satellite data provide a more accurate insight to stakeholders. While traditional data collection only provides aggregated information about regions as a whole—with statistical uncertainty due to regional limitations— earth observations can provide data at scale with high spatial granularity.

The winners of this challenge are: KarimAmer from Egypt in 1st place, youngtard from Nigeria in 2nd place and Team Be_CarEFul (Lawrence_Moruye from Kenya & Mohamed_Salam_Jedidi from Tunisia), in 3rd place.

A special thank you to the 1st and 2nd place winners for sharing some insights into how they succeeded in this challenge.

Find their solutions here.

Name: Karim Amer (1st place)

Zindi handle: KarimAmer

Where are you from? Egypt

Tell us a bit about yourself?

I am a research assistant and MSc student at Nile University. During my study, I spent one year as an intern at Siemens Healthineers in the USA.

Tell us about the approach you took.

I used a deep neural network of multiple convolutional and recurrent layers to classify the crop of input fields. The model benefits not only from the field information but also from the surroundings. To be more specific, a small patch is cropped around each field (mostly wider than field size) alongside a patch mask with ones at pixels belonging to the field and used both as inputs for the model. The first key success to make it work is to apply different augmentations including flipping, rotation, random cropping, mixup and time augmentation. The second key to success is using a Bagging Ensemble by training the model 10 times on different 85% subset of training data.

What were the things that made the difference for you that you think others can learn from?

Firstly, understanding the data and having a good local validation strategy is crucial to build a generalisable model. Secondly, using heavy augmentation is a must when training deep models on small datasets. Finally, ensemble stabilises the results in testing.

What are the biggest areas of opportunity you see in AI in Africa over the next few years?

Climate change and agriculture are very important areas to work on using AI.

What are you looking forward to most about the Zindi community?

I look forward to work with fellow Zindians to solve more real life problems using AI algorithms.

Name: Femi Sotonwa (2nd place)

Zindi handle: youngtard

Where are you from? Nigeria

Tell us a bit about yourself?

I’m a competitive Data Scientist and an accounting graduate.

Tell us about the approach you took.

My solution is an ensemble weighted average of two approaches. The first approach involved training with 3 sets of features - pixel values of each farmland; about 10 vegetation/spectral indices (e.g. NDVI, AVI etc.), and their relevant statistics; and spatial features (e.g area of farmland). The second approach involved training with only pixel values, and their relevant statistics.

What were the things that made the difference for you that you think others can learn from?

Technical - Model ensembles really helped/helps.

General - I never got tired of trying different things and testing various hypotheses up until the last days. My breakthrough on the Leaderboard that I was satisfied with came with only a day to the competition end. One can learn from this never to give up.

What are the biggest areas of opportunity you see in AI in Africa over the next few years?

Adoption of Deep Learning solutions in Computer Vision, and Natural Language Processing.

What are you looking forward to most about the Zindi community?

The growth of the community, and platform.

This competition is part of the Computer Vision for Agriculture (CV4A) Workshop at the 2020 ICLR conference and is designed and organized by Radiant Earth Foundation with support from PlantVillage in providing the ground reference data. Competition prizes are sponsored by Microsoft AI for Earth and Descartes Labs.

Watch the full Computer Vision for Agriculture (CV4A) workshop here and the presentations of our winners!

What are your thoughts on our winners' feedback? Engage via the Discussion page or leave a comment on social media.